We build AI systems for consequential analytical work.

IRIS is an AI harness for fundamental analysts.

Your models are static. The world isn't.

Import your spreadsheet and IRIS reads the model, explains what drives the forecasts and how the calculations connect, and preserves that understanding in Model Memory. AI is embedded directly in the workbook: =AI() is a formula, governed by your . As your view evolves, IRIS preserves the reasoning behind each change.

In the PM View, AI reasons across those models - their Methods, evidence, and history - so you can ask one investment question across the book and trace the answer back to each model.

Launch IRIS Request Demo

Case Study: Where does AI-related debt exposure sit?

IRIS workbook with the Research Rail beside the JPM model — Research Rail / Inside the model

When the world changes, which of my models should change with it - and why?

PM View / Across the book

Turn your models into an ecosystem.

Every model holds investment knowledge. IRIS makes that knowledge usable across the book.

PM Search uses AI to reason across the forecasts, Methods, evidence, and history inside your models. Ask one economic question across the portfolio, then follow the answer into the reasoning behind each company's view.

A move in credit spreads may matter to funding cost in one company, reinvestment yield in another, and expected losses somewhere else.

IRIS keeps those economic relationships with each model, making them durable and queryable across the book while preserving how each company responds differently.

Work done in one name becomes usable knowledge when the same driver matters elsewhere. The PM can compare assumptions, explore a shared economic scenario, and find the evidence behind a view before opening individual workbooks.

Where does the same economic driver matter across the book? How do our models respond differently to tighter credit? Which assumptions and evidence explain those differences? Where does the existing view deserve another look?

Understanding compounds across the investment process.

IRIS Models page with portfolio scope, model search, active models, and recent model activity — Models / Current corpus The book becomes searchable analytical work, not a directory of files.

IRIS comparing controlled credit-condition scenarios across models in the current corpus — PM Search / Across the book Ask one economic question across the models while preserving how each one responds differently.

The ecosystem at work / Overnight monitoring

What changed overnight?

That same connected view makes the morning review possible. Each night, IRIS refreshes market and economic data, identifies which models are affected, and reruns the analysis that depends on what changed.

By morning, PM Search can ask across the refreshed book: what changed, which models moved, which views held, and where attention belongs. Each answer leads back to the forecast, Method, evidence, and prior work behind it.

The PM sees where to look. The analyst sees why.

IRIS answering what changed last night across the current model corpus — PM Search / Morning review See what changed across the connected book and where attention belongs.

The Research Rail

Something moved. Now what?

The PM View points to a name. The Research Rail answers against the governed model state: the Model, Methods, evidence, external dependencies, Runs and Results, and saved analyst memory.

Inspect the Method, evidence, prior work, and what a changed view would affect - then ask IRIS a question, save a thought, or tell IRIS what changed without leaving the model. What matters can become a durable Saved Note or a proposed Method change for the analyst to review.

Why is this estimate here? What evidence supports this Method? Save this thought for later. I think deployment matters more than spreads here. What would this change affect in the model?

IRIS can preview a proposed change before the analyst applies it to the model.

IRIS workbook with the JPM model and Research Rail showing forecast responsibilities and items worth a closer look — Research Rail / Beside the workbook The spreadsheet stays visible while the Rail reasons against the model's governed state and saved analyst memory.

Copilot helps you work in the spreadsheet.

IRIS helps you reason with, govern, and remember the model over time.

Analyst View / Inside the model

Work the name with the model already understood.

Model Logic

Model Logic is how your model thinks.

Import a workbook and IRIS reads the model to understand how it works: what drives the forecasts, how the calculations connect, and how operating assumptions flow through to cash and funding needs.

Before frontier AI reasons about the workbook, IRIS establishes what can be known exactly: periods, formulas, dependencies, forecast structure, responsibilities, historical relationships, and the relevant inputs and outputs.

IRIS explains that understanding in plain language and preserves it in Model Memory as a durable baseline. The model can change later without rewriting what IRIS knew at import.

Methods

Frontier models give you the consensus model. IRIS turns it into yours.

The differentiated forecast comes from how the analyst weighs evidence, frames the drivers, and decides when the view should change. IRIS makes that judgment explicit as a Method - readable, editable, reviewable, and executable through the workbook. IRIS writes Methods. Analysts own them.

=AI() brings that Method into the spreadsheet as a formula. The Method governs how AI forms the forecast; the formula makes that forecast part of the model's calculations.

CoreWeave Model Imported record in Model Memory, with IRIS explaining how revenue growth, profitability, cash flow, and financing connect in the model — Model Logic / Understanding at import IRIS reads CoreWeave's model and explains how it works. That understanding remains the first record in Model Memory, even as the model evolves. Open the screenshot to read the explanation.

IRIS Method for retail same-store-sales guidance, expressed as readable analytical instructions — Method / Same-store-sales guidance The evidence hierarchy and analytical rules behind a forecast remain visible to the analyst.

Across the book / Inside the model

Case Study: Where does AI-related debt exposure sit?

Start with CoreWeave (CRWV). Trace the financing obligations, ask who bears the exposure, and examine what would change its value. Each answer sets up the next question, from the scheduled debt to the evidence needed for an investment view.

The inquiry starts with evidence already in the model: lease rows J116 and J117, the Cash Interest Rate Method at J108:O108, delayed-draw term loan (DDTL) and OEM financing disclosures, undrawn capacity at J119, and equity at J121. From there, it asks whether the exposure can be connected to other companies in the book.

A useful finding from the inquiry: in the exchange below, IRIS identifies the gap between forecasting contractual debt payments and estimating a DDTL's market value. It calls for a defined valuation subject and date, a complete cash-flow schedule, market pricing, and terms and risk evidence. The next question asks how to resolve each gap.

CoreWeave workbook beside IRIS explaining the missing Method and evidence needed to estimate a DDTL market value — 01 / Establish what the evidence supports IRIS distinguishes the model's principal and borrowing-cost forecasts from the evidence needed for a market valuation. Open the screenshot to read the response.

CoreWeave Research Rail with a follow-up drafted in the composer asking which gaps need a Method change, an analyst Note, or new external evidence — 02 / Frame the next analytical step The analyst drafts the follow-up: does each unresolved item call for a Method change, a saved Note, or new evidence? Open the screenshot to read the question.

Follow the inquiry: 11 questions from obligations to an investment thesis

Where the debt sits

Which financing obligations does this model schedule explicitly, and which does it only name from the filings without a row?
For the delayed-draw term loans, what committed, drawn and undrawn amounts does the admitted evidence establish, and as of what date?
How does the model treat the OEM and software financing, and is any of it double-counted with the disclosed debt total?
What does the model assume about lease liabilities versus lease payments, and where would a committed GPU lease show up?

Who bears it

For DDTL 3.0, who is borrower, guarantor, arranger, disclosed lender and known current holder, and which of those roles has a date attached?
Which of those lenders is a public company in this corpus, and what would be needed to connect this facility to that lender's model?

How its value changes

What drives the cash interest rate in the Method, and what happens to interest expense if SOFR moves 100 bp?
Which obligations reprice with market rates and which are fixed, and does the model separate them?
What would have to be true about revenue or utilization for the scheduled repayments to be covered from operations rather than refinancing?

Where the model is thin

What evidence is missing before this model could support a mark on the DDTL, and which of those items are in the filings the corpus already admits?
Which of the unknowns you named would a Note or an Update Method resolve, and which need new admitted evidence?

Carry the conclusion forward. A Saved Note creates a durable analyst observation attached to the current governed subject. It preserves the analyst's interpretation or conclusion for later Model Memory and reasoning.

Why an AI harness

Frontier AI gets the economics, not the spreadsheet plumbing.

A generic frontier model can reason. But hand it a workbook cold and it first has to figure out the periods, formulas, forecast structure, and which parts of the file actually matter.

Some of its intelligence is spent rediscovering facts that software can establish exactly before it reaches the investment question.

IRIS does the model archaeology once. It prepares the problem before the model call, so the reasoning can be about funding cost, unit growth, margins, credit losses, pricing, volume, or operating leverage - whatever actually drives the forecast. We call this Semantic Compression. In tests it has led to a 20x reduction in token use.

The point is not a shorter prompt. It is a better-defined question and more of the model's intelligence spent on the economics.

Let computers establish the facts. Let frontier models reason about the economics.

Model Memory

Models should remember.

Why is the model different now? A changed estimate should not erase the reason the prior estimate existed - or what IRIS understood about the model before it changed.

Model Memory is how that thinking evolves.

IRIS preserves that import-time understanding, then keeps subsequent Methods, Versions, Runs, Results, and rationale attached to the analytical changes that produced them.

Follow Versions from left to right. Open the Run beneath a change. Compare its Result with what came before and recover the reasoning that produced it. The lineage replaces overwritten cells and scattered notes with a record an analyst can actually revisit.

IRIS Model Memory lineage graph with Versions connected from left to right and the current Version selected — Version by Version, Run by Run: the analysis stays attached to the model change it produced.

Back to the analyst

Build on your investment judgment.

Excel stays the model.

Frontier AI provides the intelligence.

The analyst owns the judgment.

IRIS is the harness that keeps them connected.

Make the reasoning behind each forecast explicit, preserve it as the view evolves, and put that knowledge to work across the book.

Launch IRIS -> Read the article Request Demo